Defend against adversarial examples, prompt injection, data poisoning, model theft, unsafe tool use, insecure dependencies, and sensitive-data exposure.
